Israel is experimenting with Refaim (Unit 888), a connected warfare system combining drones, robotics, and AI (Maven, Gospel) for rapid joint strikes.
Summary
The IDF has created Refaim (Unit 888), an experimental formation designed to merge ground combat, drones, robotics, and real-time intelligence. The mission is simple and demanding: detect, qualify, and then engage the target in minutes, coordinating ground, air, and naval forces via a tight sensor-shooter loop. This unit, created in 2019 as part of the Tnufa plan, serves as a test bed for doctrines, sensors, and decision support systems. On the software side, the United States has developed Project Maven, while Israel has developed its own chain, Gospel. These algorithmic targeting tools promise faster verdicts, with mandatory human supervision to reduce errors. Public figures illustrate the scale of the shift: in February 2024, the United States struck more than 85 targets in Iraq and Syria after AI-assisted sorting; Israel claims similar gains in decision-making speed and “all-weather” capability. Behind the efficiency, however, there remain obstacles: data quality, bias prevention, decision traceability, and legal control.
The Refaim system and its experimental mission
Refaim is officially the “Multi-Dimensional Unit,” attached to the ground forces, battalion-sized, and designed since its launch in 2019 to test new combinations of sensors and effectors. It brings together light infantry, engineering, anti-tank units, armed drones, and mini-robots, with a C2 cell that prioritizes decision-making speed. This architecture responds to a very concrete need: to shorten the “spot-and-strike” chain in dense urban or semi-urban environments, where deconfliction is difficult and the window of opportunity is measured in tens of seconds. The unit came to prominence during the fighting in Gaza, including after October 7, 2023; its first commander, Colonel Roi Levy, lost his life there, proof that the experiment is taking place in the field.
In the field, the approach is modular: small teams combining drone operators, infantry group leaders, sensor specialists, and intelligence officers. A micro-network connects optical imaging, electromagnetic detection, and telemetry, then transmits an enriched track to C2. The command post then distributes missions to various vectors: precision mortars, loitering munitions, fighter jets and helicopters, and even coastal naval support. This interoperability requires a common grammar (track format, rules of engagement, prioritization) and precision standards; the typical target is metric scale for the final location. In practice, Refaim’s added value lies in its ability to assemble existing technological building blocks into a controlled and reproducible doctrine of use, rather than in a “miracle” sensor. Feedback is then shared with other brigades and manufacturers, accelerating the organic innovation loop.
The sensor-shooter loop: drones, robotics, and real-time C2
The sensor-shooter loop is based on a simple idea: compressing time. Lightweight quadcopters provide local persistence; fixed-wing drones, which have greater endurance, widen the observation cone; ground robots reduce human risk at entry points. At the center, a C2 gateway aggregates streams (video, RF, telemetry, mapping), applies detection filters, and calculates the probability of identification before assigning the most relevant effector. In a dense neighborhood, for example, a nano-UAV can confirm a thermal signature, while a loitering munition waits 3–5 km away for authorization. The savings in terms of effects are significant: the first available munition can strike within a minute, instead of mobilizing a heavier and more expensive aircraft.
Robotics offers another advantage: it “freezes” the enemy’s maneuvers. Low-cost, widely distributed sensors, multiplying the angles of view, make movement risky for the enemy. Countermeasures are still possible (jamming, decoys, saturation fire), but the redundancy of sensors and the variety of spectra (optical, thermal, RF) force the enemy to consume their own resources more quickly. The challenge, therefore, is no longer access to the image, but prioritization: which stream to watch, which lead to pursue, which risk to accept. This is precisely where Refaim comes in with its methods: sorting interface, response “playbooks,” and delegation to teams in the field to avoid overloading the command center. Recent operations have shown more compact sequences, with a tighter “detection time/neutralization time” ratio, ensuring efficiency in saturated urban areas.
Algorithmic targeting: Project Maven, Gospel, and military AI
Beyond drones, the major gain comes from algorithmic targeting. Project Maven was created on the American side to accelerate the annotation and detection of objects in massive flows. In February 2024, Washington struck more than 85 targets in Iraq and Syria, an episode often cited to illustrate the combination of “algorithmic sorting + coordinated strikes.” The tool does not “order” anything: it proposes priorities to humans who validate, reorder, or reject them. Israel has designed Gospel, and other chains such as Lavender have been publicly reported; the best description of them is as “decision support tools” rather than general-purpose AI. The tactical benefit is clear: load the algorithm with monotonous tasks (image scanning, cross-correlations) to free up the analyst for the final qualification.
Technically, three building blocks dominate: computer vision, multi-sensor fusion, and uncertainty management. Vision detects silhouettes, vehicles, antennas; fusion links these objects to SIGINT or HUMINT traces; uncertainty management assigns probabilities and displays a hierarchy. The gains are measured in minutes saved, but also in traceability: each recommendation must be audited retrospectively. This is where human supervision remains non-negotiable. Risks exist: noisy data, bias, decoy attacks. Military lawyers insist on the obligation of proportionality and the requirement for independent sources before any firing. In this approach, Refaim serves as a doctrinal testing ground: when does the algorithm make a proposal, who arbitrates, what is the minimum evidence required, what level of explainability is acceptable when the firing window is 90 seconds?
The human factor: recruitment, culture, and an old debate
Technological sophistication does not eliminate the issue of profiles. Refaim’s selection process values endurance, initiative, and the ability to reason under pressure. The cultural debate is not new: in the 2000s, articles attributed to the IDF a distrust of D&D-type role-playing gamers, on the grounds of an alleged “disconnect from reality.” These texts have fueled a persistent myth; whether or not they reflect the policy of the time, they do not describe the reality of a unit in 2025, where physical fitness is sought as much as tactical creativity and analytical rigor. Conversely, the U.S. military advocates the use of games and wargames as professional tools for training in planning and decision-making. The useful question is not “who plays,” but “who knows how to formalize an uncertain situation, make assumptions, test branches, and choose quickly.”
In Refaim’s mixed teams, a group leader must switch from operator language to analyst language in seconds: validate a lead, refuse a strike, or switch the effector to target 2 because target 1 is out of range. This cognitive plasticity can be worked on more than we think: feedback, error mapping, “hourglass” scenarios where each minute removes an option. Digital tools provide speed, but they also amplify human error if governance is lax. Hence the need for procedures, tolerable false positive quotas, and a register of case studies to feed into training. Technology does not absolve anything; it brings team discipline to light.
Operational and industrial consequences
The Refaim experiment sheds light on the short-term trajectory: scattered sensors, low-cost armed drones, and flexible C2. Purchases are shifting toward the software “stack” (ingestion, correlation, visualization, audit) and consumable sensors. Useful indicators are prosaic: median time between detection and neutralization, ratio of confirmed leads, false positive rate, time of occupation of an effector per target. Forces capable of achieving medians of less than five minutes on moving targets have a clear advantage in urban areas. Repeating these performances requires robust military AI chains and attrition logistics (batches of batteries, motors, ready-to-use sensors).
At the regional level, this integration accelerates competition: each player learns to undermine the opposing sensor ecosystem (jamming, destruction, deception) and to protect its own. For industry, demand will focus on replaceable modules, plug-and-play drones, and interoperable C2 middleware. The message is clear: value lies in the assembly—the sensor-shooter loop—rather than in an isolated drone. European armies, often fragmented by country and standards, will have to decide: adopt common C2 cores and compatible sensor catalogs, or accept lower efficiency. Feedback from Refaim serves as a guide here: modular architecture, public performance metrics, and rapid iteration between the field and the design office.
War Wings Daily is an independant magazine.